ML engineer
The one who trains a model and keeps it in service
An ML engineer takes a known learning method and makes it work on one organisation’s data, all the way to production, like an engineer building a bridge with a proven technique.
- Floor 1The Model
- P4Writes the model
- Established titleIt has been in adverts long enough to mean something.
What they answer for
The deliverableA model trained on real data, served to users, with its drift watched. If they leave, a service degrades without anything stopping.
The profile
Five activities, scored 0 to 3- ResearchProducing knowledge that does not yet exist.
- BuildShipping a system that runs, deploys and breaks.
- OperateKeeping it in production: cost, incidents, drift, on-call.
- VerifyMeasuring, testing, attacking. Producing a verdict that holds.
- LeadDeciding, persuading, driving adoption, answering to others.
What the work is
The method already exists
The difference from research fits in one sentence: the method already exists, in a paper or in a library. What remains is no simpler, but it is something else: getting the data, cleaning it, training, measuring, serving, and watching what degrades.
What actually costs
What actually costs is almost never the model. It is the data: where it came from, the right to use it, its quality, and the fact that it shifts underneath the service while the service runs. A model that loses five points in six months was not trained badly; it was trained on a world that moved.
The trap
The trap is the confusion with the AI engineer, sustained by adverts because the second title pays better. One question settles it: do you train here, or do you borrow? An advert asking for both describes two posts, and you will only hold one.
A week in the role
- Data pipelines: collection, cleaning, the split between training and test.
- Training runs, and above all the waiting they involve, during which you prepare the next one.
- Measurement: what gained, what lost, and on which population.
- Monitoring: drift shows up over weeks, never in a unit test.
Ways in
- From software development with training in machine learning, which is the commonest route.
- From data science, accepting that you answer for a service and no longer for a conclusion.
- What the role does not require, whatever the advert says: a doctorate.
Reading an advert
2 signs- The advert mentions training and large language models in the same sentence.
- Almost nobody trains a large language model. This means fine-tuning, or an AI engineer post under the wrong name.
- Nothing is said about where the data comes from.
- That is the real subject of the trade, and the first thing that stalls a project. An advert silent on it has not met the problem yet.